2D FFT / DFT + NxN Convolution (image processing) [SOLVED] [modified]

Dusan Paulovic

EDIT: I have it solved, have found a similar library as that mine and there it is implemented in Java, so I just looked for a way to do it. It is needed to use tile-overlapping, so I modified my code and now it's working... This is that page: http://www.jhlabs.com/[^] Original message: Hi all, does anybody know some good source to learn apply a fast convolution filter to image? I am looking for kernels with 20x20 an bigger and need to process image fast, so I need to learn how to do that using FFT or DFT. I like to use GDI+ in C#. Any suggestion? Thanks in idvance...

modified on Tuesday, August 30, 2011 4:23 AM

share_holder · modified on Tuesday, August 30, 2011 4:23 AM

Have you seen how does Matlab Function fft work? Source Code available and may be useful... Regards, D

MicroVirus · modified on Tuesday, August 30, 2011 4:23 AM

I'm not sure here, but isn't it more typical to either convolve with a kernel in 'normal' (image) space? A convolution is rather easy to program, so I'm guessing you could readily program a fast implementation yourself. Just know that C# (and any .NET language for that matter) isn't known for its speed. If you really want the most out of your program, you could try at C++ or another machine-compiled language.

Dusan Paulovic

I am using native C++ to rewrite working stuff, but for testing I use C#. With GDI+ it is nice that BitmapData is same structure in .Net and native GDI+, so it is easy to migrate working code from C# to native C++. Simple convolution is what I have done already, but it is too slow (even wrote in C++) compared to PhotoShop plugins, so I am searching for FFT transform of tiles... Thanks...

Dusan Paulovic

I have not so far, but I'll give it a try... But still, I would prefer any working example programmed in some .Net language or in native C / C++. Thanks...

Lost User · modified on Tuesday, August 30, 2011 4:23 AM

Andrew Kirillov made the image procing lab which he wrote an artical on here. It is quite a good read and good library to use as well (he has other articles around it) Christian Grauss wrote a series of articales youmight find usefull. The kernal topics fall under edge detection. Image Processing for Dummies with C# and GDI+ Part 3 - Edge Detection Filters[^]

Computers have been intelligent for a long time now. It just so happens that the program writers are about as effective as a room full of monkeys trying to crank out a copy of Hamlet.

Dusan Paulovic

Well, Andrew Kirillov has in his library FFT filter, but in convolution filters he does not use them and I need exactly that. Also Christion has made good stuff for beginners, but my question is little more about advanced image processing than about beginner tutorial.

Chris Losinger · modified on Tuesday, August 30, 2011 4:23 AM

if your kernels are separable, you can do the operation as a pair of 1-D convolutions (once in X, once in Y). that will change your complexity from N^2 to N.

image processing toolkits | batch image processing

Lost User

DusanDusan wrote:

Also Christion has made good stuff for beginners, but my question is little more about advanced image processing than about beginner tutorial.

DusanDusan wrote:

Hi all, does anybody know some good source to learn apply a fast convolution filter to image? I am looking for kernels with 20x20 an bigger and need to process image fast, so I need to learn how to do that using FFT or DFT.

Convulution filters and FFT are for the most part equivilant (different domains). Christian's articales are labeled beginner, but from your question I think it is exactly what you need. Why is it not advanced enough?? What details are you missing?

Computers have been intelligent for a long time now. It just so happens that the program writers are about as effective as a room full of monkeys trying to crank out a copy of Hamlet.

Dusan Paulovic

Yes, for most part are, but exactly that part, where they are not is in my interest. What I need is to understand how to forward transform image and mask, multiply them together and transform result back to 'normal' image. I have made some algorithms, but on borders of tiles I have some errors, so I am interested in any code to look for inspiration...

modified on Monday, August 29, 2011 3:29 PM

Dusan Paulovic

Some kernels are not separable, some are separable. I would like to make it using FFT, but do not know how...

Chris Losinger

maybe this will help: http://developer.download.nvidia.com/compute/cuda/2_2/sdk/website/projects/convolutionFFT2D/doc/convolutionFFT2D.pdf[^]

image processing toolkits | batch image processing

Dusan Paulovic

This is exactly what I am looking for. I have it implemented, but there is probably any error in my coding, becouse on borders of result image I have processed pixels from opposite sides of tiles. Do you know about any code that is doing that correctly? Or is it normal? Must be used overlapping of input tiles? Thanks

Chris Losinger

sorry, i don't know of any code that does this. but, that document spends a lot of time talking about how to arrange your kernel to handle the wrap-around problem.

image processing toolkits | batch image processing